diff options
Diffstat (limited to 'llvm/lib/Target/X86/MCTargetDesc')
| -rw-r--r-- | llvm/lib/Target/X86/MCTargetDesc/CMakeLists.txt | 3 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/MCTargetDesc/X86AsmBackend.cpp | 452 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp | 1055 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.cpp | 36 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.h | 11 | ||||
| -rw-r--r-- | llvm/lib/Target/X86/MCTargetDesc/X86MachObjectWriter.cpp | 554 |
6 files changed, 2111 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/MCTargetDesc/CMakeLists.txt b/llvm/lib/Target/X86/MCTargetDesc/CMakeLists.txt index ca88f8ffd08..d059aad1368 100644 --- a/llvm/lib/Target/X86/MCTargetDesc/CMakeLists.txt +++ b/llvm/lib/Target/X86/MCTargetDesc/CMakeLists.txt @@ -1,6 +1,9 @@ add_llvm_library(LLVMX86Desc + X86AsmBackend.cpp X86MCTargetDesc.cpp X86MCAsmInfo.cpp + X86MCCodeEmitter.cpp + X86MachObjectWriter.cpp ) # Hack: we need to include 'main' target directory to grab private headers diff --git a/llvm/lib/Target/X86/MCTargetDesc/X86AsmBackend.cpp b/llvm/lib/Target/X86/MCTargetDesc/X86AsmBackend.cpp new file mode 100644 index 00000000000..7e925db119a --- /dev/null +++ b/llvm/lib/Target/X86/MCTargetDesc/X86AsmBackend.cpp @@ -0,0 +1,452 @@ +//===-- X86AsmBackend.cpp - X86 Assembler Backend -------------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// + +#include "llvm/MC/TargetAsmBackend.h" +#include "MCTargetDesc/X86BaseInfo.h" +#include "MCTargetDesc/X86FixupKinds.h" +#include "llvm/ADT/Twine.h" +#include "llvm/MC/MCAssembler.h" +#include "llvm/MC/MCELFObjectWriter.h" +#include "llvm/MC/MCExpr.h" +#include "llvm/MC/MCFixupKindInfo.h" +#include "llvm/MC/MCMachObjectWriter.h" +#include "llvm/MC/MCObjectWriter.h" +#include "llvm/MC/MCSectionCOFF.h" +#include "llvm/MC/MCSectionELF.h" +#include "llvm/MC/MCSectionMachO.h" +#include "llvm/Object/MachOFormat.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/ELF.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/raw_ostream.h" +#include "llvm/Target/TargetRegistry.h" +using namespace llvm; + +// Option to allow disabling arithmetic relaxation to workaround PR9807, which +// is useful when running bitwise comparison experiments on Darwin. We should be +// able to remove this once PR9807 is resolved. +static cl::opt<bool> +MCDisableArithRelaxation("mc-x86-disable-arith-relaxation", + cl::desc("Disable relaxation of arithmetic instruction for X86")); + +static unsigned getFixupKindLog2Size(unsigned Kind) { + switch (Kind) { + default: assert(0 && "invalid fixup kind!"); + case FK_PCRel_1: + case FK_Data_1: return 0; + case FK_PCRel_2: + case FK_Data_2: return 1; + case FK_PCRel_4: + case X86::reloc_riprel_4byte: + case X86::reloc_riprel_4byte_movq_load: + case X86::reloc_signed_4byte: + case X86::reloc_global_offset_table: + case FK_Data_4: return 2; + case FK_PCRel_8: + case FK_Data_8: return 3; + } +} + +namespace { + +class X86ELFObjectWriter : public MCELFObjectTargetWriter { +public: + X86ELFObjectWriter(bool is64Bit, Triple::OSType OSType, uint16_t EMachine, + bool HasRelocationAddend) + : MCELFObjectTargetWriter(is64Bit, OSType, EMachine, HasRelocationAddend) {} +}; + +class X86AsmBackend : public TargetAsmBackend { +public: + X86AsmBackend(const Target &T) + : TargetAsmBackend() {} + + unsigned getNumFixupKinds() const { + return X86::NumTargetFixupKinds; + } + + const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const { + const static MCFixupKindInfo Infos[X86::NumTargetFixupKinds] = { + { "reloc_riprel_4byte", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel }, + { "reloc_riprel_4byte_movq_load", 0, 4 * 8, MCFixupKindInfo::FKF_IsPCRel}, + { "reloc_signed_4byte", 0, 4 * 8, 0}, + { "reloc_global_offset_table", 0, 4 * 8, 0} + }; + + if (Kind < FirstTargetFixupKind) + return TargetAsmBackend::getFixupKindInfo(Kind); + + assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() && + "Invalid kind!"); + return Infos[Kind - FirstTargetFixupKind]; + } + + void ApplyFixup(const MCFixup &Fixup, char *Data, unsigned DataSize, + uint64_t Value) const { + unsigned Size = 1 << getFixupKindLog2Size(Fixup.getKind()); + + assert(Fixup.getOffset() + Size <= DataSize && + "Invalid fixup offset!"); + for (unsigned i = 0; i != Size; ++i) + Data[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8)); + } + + bool MayNeedRelaxation(const MCInst &Inst) const; + + void RelaxInstruction(const MCInst &Inst, MCInst &Res) const; + + bool WriteNopData(uint64_t Count, MCObjectWriter *OW) const; +}; +} // end anonymous namespace + +static unsigned getRelaxedOpcodeBranch(unsigned Op) { + switch (Op) { + default: + return Op; + + case X86::JAE_1: return X86::JAE_4; + case X86::JA_1: return X86::JA_4; + case X86::JBE_1: return X86::JBE_4; + case X86::JB_1: return X86::JB_4; + case X86::JE_1: return X86::JE_4; + case X86::JGE_1: return X86::JGE_4; + case X86::JG_1: return X86::JG_4; + case X86::JLE_1: return X86::JLE_4; + case X86::JL_1: return X86::JL_4; + case X86::JMP_1: return X86::JMP_4; + case X86::JNE_1: return X86::JNE_4; + case X86::JNO_1: return X86::JNO_4; + case X86::JNP_1: return X86::JNP_4; + case X86::JNS_1: return X86::JNS_4; + case X86::JO_1: return X86::JO_4; + case X86::JP_1: return X86::JP_4; + case X86::JS_1: return X86::JS_4; + } +} + +static unsigned getRelaxedOpcodeArith(unsigned Op) { + switch (Op) { + default: + return Op; + + // IMUL + case X86::IMUL16rri8: return X86::IMUL16rri; + case X86::IMUL16rmi8: return X86::IMUL16rmi; + case X86::IMUL32rri8: return X86::IMUL32rri; + case X86::IMUL32rmi8: return X86::IMUL32rmi; + case X86::IMUL64rri8: return X86::IMUL64rri32; + case X86::IMUL64rmi8: return X86::IMUL64rmi32; + + // AND + case X86::AND16ri8: return X86::AND16ri; + case X86::AND16mi8: return X86::AND16mi; + case X86::AND32ri8: return X86::AND32ri; + case X86::AND32mi8: return X86::AND32mi; + case X86::AND64ri8: return X86::AND64ri32; + case X86::AND64mi8: return X86::AND64mi32; + + // OR + case X86::OR16ri8: return X86::OR16ri; + case X86::OR16mi8: return X86::OR16mi; + case X86::OR32ri8: return X86::OR32ri; + case X86::OR32mi8: return X86::OR32mi; + case X86::OR64ri8: return X86::OR64ri32; + case X86::OR64mi8: return X86::OR64mi32; + + // XOR + case X86::XOR16ri8: return X86::XOR16ri; + case X86::XOR16mi8: return X86::XOR16mi; + case X86::XOR32ri8: return X86::XOR32ri; + case X86::XOR32mi8: return X86::XOR32mi; + case X86::XOR64ri8: return X86::XOR64ri32; + case X86::XOR64mi8: return X86::XOR64mi32; + + // ADD + case X86::ADD16ri8: return X86::ADD16ri; + case X86::ADD16mi8: return X86::ADD16mi; + case X86::ADD32ri8: return X86::ADD32ri; + case X86::ADD32mi8: return X86::ADD32mi; + case X86::ADD64ri8: return X86::ADD64ri32; + case X86::ADD64mi8: return X86::ADD64mi32; + + // SUB + case X86::SUB16ri8: return X86::SUB16ri; + case X86::SUB16mi8: return X86::SUB16mi; + case X86::SUB32ri8: return X86::SUB32ri; + case X86::SUB32mi8: return X86::SUB32mi; + case X86::SUB64ri8: return X86::SUB64ri32; + case X86::SUB64mi8: return X86::SUB64mi32; + + // CMP + case X86::CMP16ri8: return X86::CMP16ri; + case X86::CMP16mi8: return X86::CMP16mi; + case X86::CMP32ri8: return X86::CMP32ri; + case X86::CMP32mi8: return X86::CMP32mi; + case X86::CMP64ri8: return X86::CMP64ri32; + case X86::CMP64mi8: return X86::CMP64mi32; + + // PUSH + case X86::PUSHi8: return X86::PUSHi32; + case X86::PUSHi16: return X86::PUSHi32; + case X86::PUSH64i8: return X86::PUSH64i32; + case X86::PUSH64i16: return X86::PUSH64i32; + } +} + +static unsigned getRelaxedOpcode(unsigned Op) { + unsigned R = getRelaxedOpcodeArith(Op); + if (R != Op) + return R; + return getRelaxedOpcodeBranch(Op); +} + +bool X86AsmBackend::MayNeedRelaxation(const MCInst &Inst) const { + // Branches can always be relaxed. + if (getRelaxedOpcodeBranch(Inst.getOpcode()) != Inst.getOpcode()) + return true; + + if (MCDisableArithRelaxation) + return false; + + // Check if this instruction is ever relaxable. + if (getRelaxedOpcodeArith(Inst.getOpcode()) == Inst.getOpcode()) + return false; + + + // Check if it has an expression and is not RIP relative. + bool hasExp = false; + bool hasRIP = false; + for (unsigned i = 0; i < Inst.getNumOperands(); ++i) { + const MCOperand &Op = Inst.getOperand(i); + if (Op.isExpr()) + hasExp = true; + + if (Op.isReg() && Op.getReg() == X86::RIP) + hasRIP = true; + } + + // FIXME: Why exactly do we need the !hasRIP? Is it just a limitation on + // how we do relaxations? + return hasExp && !hasRIP; +} + +// FIXME: Can tblgen help at all here to verify there aren't other instructions +// we can relax? +void X86AsmBackend::RelaxInstruction(const MCInst &Inst, MCInst &Res) const { + // The only relaxations X86 does is from a 1byte pcrel to a 4byte pcrel. + unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode()); + + if (RelaxedOp == Inst.getOpcode()) { + SmallString<256> Tmp; + raw_svector_ostream OS(Tmp); + Inst.dump_pretty(OS); + OS << "\n"; + report_fatal_error("unexpected instruction to relax: " + OS.str()); + } + + Res = Inst; + Res.setOpcode(RelaxedOp); +} + +/// WriteNopData - Write optimal nops to the output file for the \arg Count +/// bytes. This returns the number of bytes written. It may return 0 if +/// the \arg Count is more than the maximum optimal nops. +bool X86AsmBackend::WriteNopData(uint64_t Count, MCObjectWriter *OW) const { + static const uint8_t Nops[10][10] = { + // nop + {0x90}, + // xchg %ax,%ax + {0x66, 0x90}, + // nopl (%[re]ax) + {0x0f, 0x1f, 0x00}, + // nopl 0(%[re]ax) + {0x0f, 0x1f, 0x40, 0x00}, + // nopl 0(%[re]ax,%[re]ax,1) + {0x0f, 0x1f, 0x44, 0x00, 0x00}, + // nopw 0(%[re]ax,%[re]ax,1) + {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00}, + // nopl 0L(%[re]ax) + {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00}, + // nopl 0L(%[re]ax,%[re]ax,1) + {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}, + // nopw 0L(%[re]ax,%[re]ax,1) + {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}, + // nopw %cs:0L(%[re]ax,%[re]ax,1) + {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}, + }; + + // Write an optimal sequence for the first 15 bytes. + const uint64_t OptimalCount = (Count < 16) ? Count : 15; + const uint64_t Prefixes = OptimalCount <= 10 ? 0 : OptimalCount - 10; + for (uint64_t i = 0, e = Prefixes; i != e; i++) + OW->Write8(0x66); + const uint64_t Rest = OptimalCount - Prefixes; + for (uint64_t i = 0, e = Rest; i != e; i++) + OW->Write8(Nops[Rest - 1][i]); + + // Finish with single byte nops. + for (uint64_t i = OptimalCount, e = Count; i != e; ++i) + OW->Write8(0x90); + + return true; +} + +/* *** */ + +namespace { +class ELFX86AsmBackend : public X86AsmBackend { +public: + Triple::OSType OSType; + ELFX86AsmBackend(const Target &T, Triple::OSType _OSType) + : X86AsmBackend(T), OSType(_OSType) { + HasReliableSymbolDifference = true; + } + + virtual bool doesSectionRequireSymbols(const MCSection &Section) const { + const MCSectionELF &ES = static_cast<const MCSectionELF&>(Section); + return ES.getFlags() & ELF::SHF_MERGE; + } +}; + +class ELFX86_32AsmBackend : public ELFX86AsmBackend { +public: + ELFX86_32AsmBackend(const Target &T, Triple::OSType OSType) + : ELFX86AsmBackend(T, OSType) {} + + MCObjectWriter *createObjectWriter(raw_ostream &OS) const { + return createELFObjectWriter(createELFObjectTargetWriter(), + OS, /*IsLittleEndian*/ true); + } + + MCELFObjectTargetWriter *createELFObjectTargetWriter() const { + return new X86ELFObjectWriter(false, OSType, ELF::EM_386, false); + } +}; + +class ELFX86_64AsmBackend : public ELFX86AsmBackend { +public: + ELFX86_64AsmBackend(const Target &T, Triple::OSType OSType) + : ELFX86AsmBackend(T, OSType) {} + + MCObjectWriter *createObjectWriter(raw_ostream &OS) const { + return createELFObjectWriter(createELFObjectTargetWriter(), + OS, /*IsLittleEndian*/ true); + } + + MCELFObjectTargetWriter *createELFObjectTargetWriter() const { + return new X86ELFObjectWriter(true, OSType, ELF::EM_X86_64, true); + } +}; + +class WindowsX86AsmBackend : public X86AsmBackend { + bool Is64Bit; + +public: + WindowsX86AsmBackend(const Target &T, bool is64Bit) + : X86AsmBackend(T) + , Is64Bit(is64Bit) { + } + + MCObjectWriter *createObjectWriter(raw_ostream &OS) const { + return createWinCOFFObjectWriter(OS, Is64Bit); + } +}; + +class DarwinX86AsmBackend : public X86AsmBackend { +public: + DarwinX86AsmBackend(const Target &T) + : X86AsmBackend(T) { } +}; + +class DarwinX86_32AsmBackend : public DarwinX86AsmBackend { +public: + DarwinX86_32AsmBackend(const Target &T) + : DarwinX86AsmBackend(T) {} + + MCObjectWriter *createObjectWriter(raw_ostream &OS) const { + return createX86MachObjectWriter(OS, /*Is64Bit=*/false, + object::mach::CTM_i386, + object::mach::CSX86_ALL); + } +}; + +class DarwinX86_64AsmBackend : public DarwinX86AsmBackend { +public: + DarwinX86_64AsmBackend(const Target &T) + : DarwinX86AsmBackend(T) { + HasReliableSymbolDifference = true; + } + + MCObjectWriter *createObjectWriter(raw_ostream &OS) const { + return createX86MachObjectWriter(OS, /*Is64Bit=*/true, + object::mach::CTM_x86_64, + object::mach::CSX86_ALL); + } + + virtual bool doesSectionRequireSymbols(const MCSection &Section) const { + // Temporary labels in the string literals sections require symbols. The + // issue is that the x86_64 relocation format does not allow symbol + + // offset, and so the linker does not have enough information to resolve the + // access to the appropriate atom unless an external relocation is used. For + // non-cstring sections, we expect the compiler to use a non-temporary label + // for anything that could have an addend pointing outside the symbol. + // + // See <rdar://problem/4765733>. + const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section); + return SMO.getType() == MCSectionMachO::S_CSTRING_LITERALS; + } + + virtual bool isSectionAtomizable(const MCSection &Section) const { + const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section); + // Fixed sized data sections are uniqued, they cannot be diced into atoms. + switch (SMO.getType()) { + default: + return true; + + case MCSectionMachO::S_4BYTE_LITERALS: + case MCSectionMachO::S_8BYTE_LITERALS: + case MCSectionMachO::S_16BYTE_LITERALS: + case MCSectionMachO::S_LITERAL_POINTERS: + case MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS: + case MCSectionMachO::S_LAZY_SYMBOL_POINTERS: + case MCSectionMachO::S_MOD_INIT_FUNC_POINTERS: + case MCSectionMachO::S_MOD_TERM_FUNC_POINTERS: + case MCSectionMachO::S_INTERPOSING: + return false; + } + } +}; + +} // end anonymous namespace + +TargetAsmBackend *llvm::createX86_32AsmBackend(const Target &T, + const std::string &TT) { + Triple TheTriple(TT); + + if (TheTriple.isOSDarwin() || TheTriple.getEnvironment() == Triple::MachO) + return new DarwinX86_32AsmBackend(T); + + if (TheTriple.isOSWindows()) + return new WindowsX86AsmBackend(T, false); + + return new ELFX86_32AsmBackend(T, TheTriple.getOS()); +} + +TargetAsmBackend *llvm::createX86_64AsmBackend(const Target &T, + const std::string &TT) { + Triple TheTriple(TT); + + if (TheTriple.isOSDarwin() || TheTriple.getEnvironment() == Triple::MachO) + return new DarwinX86_64AsmBackend(T); + + if (TheTriple.isOSWindows()) + return new WindowsX86AsmBackend(T, true); + + return new ELFX86_64AsmBackend(T, TheTriple.getOS()); +} diff --git a/llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp b/llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp new file mode 100644 index 00000000000..7a56e5c5a29 --- /dev/null +++ b/llvm/lib/Target/X86/MCTargetDesc/X86MCCodeEmitter.cpp @@ -0,0 +1,1055 @@ +//===-- X86/X86MCCodeEmitter.cpp - Convert X86 code to machine code -------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file implements the X86MCCodeEmitter class. +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "mccodeemitter" +#include "MCTargetDesc/X86MCTargetDesc.h" +#include "MCTargetDesc/X86BaseInfo.h" +#include "MCTargetDesc/X86FixupKinds.h" +#include "llvm/MC/MCCodeEmitter.h" +#include "llvm/MC/MCExpr.h" +#include "llvm/MC/MCInst.h" +#include "llvm/MC/MCInstrInfo.h" +#include "llvm/MC/MCRegisterInfo.h" +#include "llvm/MC/MCSubtargetInfo.h" +#include "llvm/MC/MCSymbol.h" +#include "llvm/Support/raw_ostream.h" + +using namespace llvm; + +namespace { +class X86MCCodeEmitter : public MCCodeEmitter { + X86MCCodeEmitter(const X86MCCodeEmitter &); // DO NOT IMPLEMENT + void operator=(const X86MCCodeEmitter &); // DO NOT IMPLEMENT + const MCInstrInfo &MCII; + const MCSubtargetInfo &STI; + MCContext &Ctx; +public: + X86MCCodeEmitter(const MCInstrInfo &mcii, const MCSubtargetInfo &sti, + MCContext &ctx) + : MCII(mcii), STI(sti), Ctx(ctx) { + } + + ~X86MCCodeEmitter() {} + + bool is64BitMode() const { + // FIXME: Can tablegen auto-generate this? + return (STI.getFeatureBits() & X86::Mode64Bit) != 0; + } + + static unsigned GetX86RegNum(const MCOperand &MO) { + return X86_MC::getX86RegNum(MO.getReg()); + } + + // On regular x86, both XMM0-XMM7 and XMM8-XMM15 are encoded in the range + // 0-7 and the difference between the 2 groups is given by the REX prefix. + // In the VEX prefix, registers are seen sequencially from 0-15 and encoded + // in 1's complement form, example: + // + // ModRM field => XMM9 => 1 + // VEX.VVVV => XMM9 => ~9 + // + // See table 4-35 of Intel AVX Programming Reference for details. + static unsigned char getVEXRegisterEncoding(const MCInst &MI, + unsigned OpNum) { + unsigned SrcReg = MI.getOperand(OpNum).getReg(); + unsigned SrcRegNum = GetX86RegNum(MI.getOperand(OpNum)); + if ((SrcReg >= X86::XMM8 && SrcReg <= X86::XMM15) || + (SrcReg >= X86::YMM8 && SrcReg <= X86::YMM15)) + SrcRegNum += 8; + + // The registers represented through VEX_VVVV should + // be encoded in 1's complement form. + return (~SrcRegNum) & 0xf; + } + + void EmitByte(unsigned char C, unsigned &CurByte, raw_ostream &OS) const { + OS << (char)C; + ++CurByte; + } + + void EmitConstant(uint64_t Val, unsigned Size, unsigned &CurByte, + raw_ostream &OS) const { + // Output the constant in little endian byte order. + for (unsigned i = 0; i != Size; ++i) { + EmitByte(Val & 255, CurByte, OS); + Val >>= 8; + } + } + + void EmitImmediate(const MCOperand &Disp, + unsigned ImmSize, MCFixupKind FixupKind, + unsigned &CurByte, raw_ostream &OS, + SmallVectorImpl<MCFixup> &Fixups, + int ImmOffset = 0) const; + + inline static unsigned char ModRMByte(unsigned Mod, unsigned RegOpcode, + unsigned RM) { + assert(Mod < 4 && RegOpcode < 8 && RM < 8 && "ModRM Fields out of range!"); + return RM | (RegOpcode << 3) | (Mod << 6); + } + + void EmitRegModRMByte(const MCOperand &ModRMReg, unsigned RegOpcodeFld, + unsigned &CurByte, raw_ostream &OS) const { + EmitByte(ModRMByte(3, RegOpcodeFld, GetX86RegNum(ModRMReg)), CurByte, OS); + } + + void EmitSIBByte(unsigned SS, unsigned Index, unsigned Base, + unsigned &CurByte, raw_ostream &OS) const { + // SIB byte is in the same format as the ModRMByte. + EmitByte(ModRMByte(SS, Index, Base), CurByte, OS); + } + + + void EmitMemModRMByte(const MCInst &MI, unsigned Op, + unsigned RegOpcodeField, + uint64_t TSFlags, unsigned &CurByte, raw_ostream &OS, + SmallVectorImpl<MCFixup> &Fixups) const; + + void EncodeInstruction(const MCInst &MI, raw_ostream &OS, + SmallVectorImpl<MCFixup> &Fixups) const; + + void EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand, + const MCInst &MI, const MCInstrDesc &Desc, + raw_ostream &OS) const; + + void EmitSegmentOverridePrefix(uint64_t TSFlags, unsigned &CurByte, + int MemOperand, const MCInst &MI, + raw_ostream &OS) const; + + void EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, int MemOperand, + const MCInst &MI, const MCInstrDesc &Desc, + raw_ostream &OS) const; +}; + +} // end anonymous namespace + + +MCCodeEmitter *llvm::createX86MCCodeEmitter(const MCInstrInfo &MCII, + const MCSubtargetInfo &STI, + MCContext &Ctx) { + return new X86MCCodeEmitter(MCII, STI, Ctx); +} + +/// isDisp8 - Return true if this signed displacement fits in a 8-bit +/// sign-extended field. +static bool isDisp8(int Value) { + return Value == (signed char)Value; +} + +/// getImmFixupKind - Return the appropriate fixup kind to use for an immediate +/// in an instruction with the specified TSFlags. +static MCFixupKind getImmFixupKind(uint64_t TSFlags) { + unsigned Size = X86II::getSizeOfImm(TSFlags); + bool isPCRel = X86II::isImmPCRel(TSFlags); + + return MCFixup::getKindForSize(Size, isPCRel); +} + +namespace llvm { + // FIXME: TableGen this? + extern MCRegisterClass X86MCRegisterClasses[]; // In X86GenRegisterInfo.inc. +} + +/// Is32BitMemOperand - Return true if the specified instruction with a memory +/// operand should emit the 0x67 prefix byte in 64-bit mode due to a 32-bit +/// memory operand. Op specifies the operand # of the memoperand. +static bool Is32BitMemOperand(const MCInst &MI, unsigned Op) { + const MCOperand &BaseReg = MI.getOperand(Op+X86::AddrBaseReg); + const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg); + + if ((BaseReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR32RegClassID].contains(BaseReg.getReg())) || + (IndexReg.getReg() != 0 && + X86MCRegisterClasses[X86::GR32RegClassID].contains(IndexReg.getReg()))) + return true; + return false; +} + +/// StartsWithGlobalOffsetTable - Return true for the simple cases where this +/// expression starts with _GLOBAL_OFFSET_TABLE_. This is a needed to support +/// PIC on ELF i386 as that symbol is magic. We check only simple case that +/// are know to be used: _GLOBAL_OFFSET_TABLE_ by itself or at the start +/// of a binary expression. +static bool StartsWithGlobalOffsetTable(const MCExpr *Expr) { + if (Expr->getKind() == MCExpr::Binary) { + const MCBinaryExpr *BE = static_cast<const MCBinaryExpr *>(Expr); + Expr = BE->getLHS(); + } + + if (Expr->getKind() != MCExpr::SymbolRef) + return false; + + const MCSymbolRefExpr *Ref = static_cast<const MCSymbolRefExpr*>(Expr); + const MCSymbol &S = Ref->getSymbol(); + return S.getName() == "_GLOBAL_OFFSET_TABLE_"; +} + +void X86MCCodeEmitter:: +EmitImmediate(const MCOperand &DispOp, unsigned Size, MCFixupKind FixupKind, + unsigned &CurByte, raw_ostream &OS, + SmallVectorImpl<MCFixup> &Fixups, int ImmOffset) const { + const MCExpr *Expr = NULL; + if (DispOp.isImm()) { + // If this is a simple integer displacement that doesn't require a relocation, + // emit it now. + if (FixupKind != FK_PCRel_1 && + FixupKind != FK_PCRel_2 && + FixupKind != FK_PCRel_4) { + EmitConstant(DispOp.getImm()+ImmOffset, Size, CurByte, OS); + return; + } + Expr = MCConstantExpr::Create(DispOp.getImm(), Ctx); + } else { + Expr = DispOp.getExpr(); + } + + // If we have an immoffset, add it to the expression. + if ((FixupKind == FK_Data_4 || + FixupKind == MCFixupKind(X86::reloc_signed_4byte)) && + StartsWithGlobalOffsetTable(Expr)) { + assert(ImmOffset == 0); + + FixupKind = MCFixupKind(X86::reloc_global_offset_table); + ImmOffset = CurByte; + } + + // If the fixup is pc-relative, we need to bias the value to be relative to + // the start of the field, not the end of the field. + if (FixupKind == FK_PCRel_4 || + FixupKind == MCFixupKind(X86::reloc_riprel_4byte) || + FixupKind == MCFixupKind(X86::reloc_riprel_4byte_movq_load)) + ImmOffset -= 4; + if (FixupKind == FK_PCRel_2) + ImmOffset -= 2; + if (FixupKind == FK_PCRel_1) + ImmOffset -= 1; + + if (ImmOffset) + Expr = MCBinaryExpr::CreateAdd(Expr, MCConstantExpr::Create(ImmOffset, Ctx), + Ctx); + + // Emit a symbolic constant as a fixup and 4 zeros. + Fixups.push_back(MCFixup::Create(CurByte, Expr, FixupKind)); + EmitConstant(0, Size, CurByte, OS); +} + +void X86MCCodeEmitter::EmitMemModRMByte(const MCInst &MI, unsigned Op, + unsigned RegOpcodeField, + uint64_t TSFlags, unsigned &CurByte, + raw_ostream &OS, + SmallVectorImpl<MCFixup> &Fixups) const{ + const MCOperand &Disp = MI.getOperand(Op+X86::AddrDisp); + const MCOperand &Base = MI.getOperand(Op+X86::AddrBaseReg); + const MCOperand &Scale = MI.getOperand(Op+X86::AddrScaleAmt); + const MCOperand &IndexReg = MI.getOperand(Op+X86::AddrIndexReg); + unsigned BaseReg = Base.getReg(); + + // Handle %rip relative addressing. + if (BaseReg == X86::RIP) { // [disp32+RIP] in X86-64 mode + assert(is64BitMode() && "Rip-relative addressing requires 64-bit mode"); + assert(IndexReg.getReg() == 0 && "Invalid rip-relative address"); + EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS); + + unsigned FixupKind = X86::reloc_riprel_4byte; + + // movq loads are handled with a special relocation form which allows the + // linker to eliminate some loads for GOT references which end up in the + // same linkage unit. + if (MI.getOpcode() == X86::MOV64rm) + FixupKind = X86::reloc_riprel_4byte_movq_load; + + // rip-relative addressing is actually relative to the *next* instruction. + // Since an immediate can follow the mod/rm byte for an instruction, this + // means that we need to bias the immediate field of the instruction with + // the size of the immediate field. If we have this case, add it into the + // expression to emit. + int ImmSize = X86II::hasImm(TSFlags) ? X86II::getSizeOfImm(TSFlags) : 0; + + EmitImmediate(Disp, 4, MCFixupKind(FixupKind), + CurByte, OS, Fixups, -ImmSize); + return; + } + + unsigned BaseRegNo = BaseReg ? GetX86RegNum(Base) : -1U; + + // Determine whether a SIB byte is needed. + // If no BaseReg, issue a RIP relative instruction only if the MCE can + // resolve addresses on-the-fly, otherwise use SIB (Intel Manual 2A, table + // 2-7) and absolute references. + + if (// The SIB byte must be used if there is an index register. + IndexReg.getReg() == 0 && + // The SIB byte must be used if the base is ESP/RSP/R12, all of which + // encode to an R/M value of 4, which indicates that a SIB byte is + // present. + BaseRegNo != N86::ESP && + // If there is no base register and we're in 64-bit mode, we need a SIB + // byte to emit an addr that is just 'disp32' (the non-RIP relative form). + (!is64BitMode() || BaseReg != 0)) { + + if (BaseReg == 0) { // [disp32] in X86-32 mode + EmitByte(ModRMByte(0, RegOpcodeField, 5), CurByte, OS); + EmitImmediate(Disp, 4, FK_Data_4, CurByte, OS, Fixups); + return; + } + + // If the base is not EBP/ESP and there is no displacement, use simple + // indirect register encoding, this handles addresses like [EAX]. The + // encoding for [EBP] with no displacement means [disp32] so we handle it + // by emitting a displacement of 0 below. + if (Disp.isImm() && Disp.getImm() == 0 && BaseRegNo != N86::EBP) { + EmitByte(ModRMByte(0, RegOpcodeField, BaseRegNo), CurByte, OS); + return; + } + + // Otherwise, if the displacement fits in a byte, encode as [REG+disp8]. + if (Disp.isImm() && isDisp8(Disp.getImm())) { + EmitByte(ModRMByte(1, RegOpcodeField, BaseRegNo), CurByte, OS); + EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups); + return; + } + + // Otherwise, emit the most general non-SIB encoding: [REG+disp32] + EmitByte(ModRMByte(2, RegOpcodeField, BaseRegNo), CurByte, OS); + EmitImmediate(Disp, 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS, + Fixups); + return; + } + + // We need a SIB byte, so start by outputting the ModR/M byte first + assert(IndexReg.getReg() != X86::ESP && + IndexReg.getReg() != X86::RSP && "Cannot use ESP as index reg!"); + + bool ForceDisp32 = false; + bool ForceDisp8 = false; + if (BaseReg == 0) { + // If there is no base register, we emit the special case SIB byte with + // MOD=0, BASE=5, to JUST get the index, scale, and displacement. + EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS); + ForceDisp32 = true; + } else if (!Disp.isImm()) { + // Emit the normal disp32 encoding. + EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS); + ForceDisp32 = true; + } else if (Disp.getImm() == 0 && + // Base reg can't be anything that ends up with '5' as the base + // reg, it is the magic [*] nomenclature that indicates no base. + BaseRegNo != N86::EBP) { + // Emit no displacement ModR/M byte + EmitByte(ModRMByte(0, RegOpcodeField, 4), CurByte, OS); + } else if (isDisp8(Disp.getImm())) { + // Emit the disp8 encoding. + EmitByte(ModRMByte(1, RegOpcodeField, 4), CurByte, OS); + ForceDisp8 = true; // Make sure to force 8 bit disp if Base=EBP + } else { + // Emit the normal disp32 encoding. + EmitByte(ModRMByte(2, RegOpcodeField, 4), CurByte, OS); + } + + // Calculate what the SS field value should be... + static const unsigned SSTable[] = { ~0, 0, 1, ~0, 2, ~0, ~0, ~0, 3 }; + unsigned SS = SSTable[Scale.getImm()]; + + if (BaseReg == 0) { + // Handle the SIB byte for the case where there is no base, see Intel + // Manual 2A, table 2-7. The displacement has already been output. + unsigned IndexRegNo; + if (IndexReg.getReg()) + IndexRegNo = GetX86RegNum(IndexReg); + else // Examples: [ESP+1*<noreg>+4] or [scaled idx]+disp32 (MOD=0,BASE=5) + IndexRegNo = 4; + EmitSIBByte(SS, IndexRegNo, 5, CurByte, OS); + } else { + unsigned IndexRegNo; + if (IndexReg.getReg()) + IndexRegNo = GetX86RegNum(IndexReg); + else + IndexRegNo = 4; // For example [ESP+1*<noreg>+4] + EmitSIBByte(SS, IndexRegNo, GetX86RegNum(Base), CurByte, OS); + } + + // Do we need to output a displacement? + if (ForceDisp8) + EmitImmediate(Disp, 1, FK_Data_1, CurByte, OS, Fixups); + else if (ForceDisp32 || Disp.getImm() != 0) + EmitImmediate(Disp, 4, MCFixupKind(X86::reloc_signed_4byte), CurByte, OS, + Fixups); +} + +/// EmitVEXOpcodePrefix - AVX instructions are encoded using a opcode prefix +/// called VEX. +void X86MCCodeEmitter::EmitVEXOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, + int MemOperand, const MCInst &MI, + const MCInstrDesc &Desc, + raw_ostream &OS) const { + bool HasVEX_4V = false; + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_4V) + HasVEX_4V = true; + + // VEX_R: opcode externsion equivalent to REX.R in + // 1's complement (inverted) form + // + // 1: Same as REX_R=0 (must be 1 in 32-bit mode) + // 0: Same as REX_R=1 (64 bit mode only) + // + unsigned char VEX_R = 0x1; + + // VEX_X: equivalent to REX.X, only used when a + // register is used for index in SIB Byte. + // + // 1: Same as REX.X=0 (must be 1 in 32-bit mode) + // 0: Same as REX.X=1 (64-bit mode only) + unsigned char VEX_X = 0x1; + + // VEX_B: + // + // 1: Same as REX_B=0 (ignored in 32-bit mode) + // 0: Same as REX_B=1 (64 bit mode only) + // + unsigned char VEX_B = 0x1; + + // VEX_W: opcode specific (use like REX.W, or used for + // opcode extension, or ignored, depending on the opcode byte) + unsigned char VEX_W = 0; + + // VEX_5M (VEX m-mmmmm field): + // + // 0b00000: Reserved for future use + // 0b00001: implied 0F leading opcode + // 0b00010: implied 0F 38 leading opcode bytes + // 0b00011: implied 0F 3A leading opcode bytes + // 0b00100-0b11111: Reserved for future use + // + unsigned char VEX_5M = 0x1; + + // VEX_4V (VEX vvvv field): a register specifier + // (in 1's complement form) or 1111 if unused. + unsigned char VEX_4V = 0xf; + + // VEX_L (Vector Length): + // + // 0: scalar or 128-bit vector + // 1: 256-bit vector + // + unsigned char VEX_L = 0; + + // VEX_PP: opcode extension providing equivalent + // functionality of a SIMD prefix + // + // 0b00: None + // 0b01: 66 + // 0b10: F3 + // 0b11: F2 + // + unsigned char VEX_PP = 0; + + // Encode the operand size opcode prefix as needed. + if (TSFlags & X86II::OpSize) + VEX_PP = 0x01; + + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_W) + VEX_W = 1; + + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_L) + VEX_L = 1; + + switch (TSFlags & X86II::Op0Mask) { + default: assert(0 && "Invalid prefix!"); + case X86II::T8: // 0F 38 + VEX_5M = 0x2; + break; + case X86II::TA: // 0F 3A + VEX_5M = 0x3; + break; + case X86II::TF: // F2 0F 38 + VEX_PP = 0x3; + VEX_5M = 0x2; + break; + case X86II::XS: // F3 0F + VEX_PP = 0x2; + break; + case X86II::XD: // F2 0F + VEX_PP = 0x3; + break; + case X86II::A6: // Bypass: Not used by VEX + case X86II::A7: // Bypass: Not used by VEX + case X86II::TB: // Bypass: Not used by VEX + case 0: + break; // No prefix! + } + + // Set the vector length to 256-bit if YMM0-YMM15 is used + for (unsigned i = 0; i != MI.getNumOperands(); ++i) { + if (!MI.getOperand(i).isReg()) + continue; + unsigned SrcReg = MI.getOperand(i).getReg(); + if (SrcReg >= X86::YMM0 && SrcReg <= X86::YMM15) + VEX_L = 1; + } + + unsigned NumOps = MI.getNumOperands(); + unsigned CurOp = 0; + bool IsDestMem = false; + + switch (TSFlags & X86II::FormMask) { + case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!"); + case X86II::MRMDestMem: + IsDestMem = true; + // The important info for the VEX prefix is never beyond the address + // registers. Don't check beyond that. + NumOps = CurOp = X86::AddrNumOperands; + case X86II::MRM0m: case X86II::MRM1m: + case X86II::MRM2m: case X86II::MRM3m: + case X86II::MRM4m: case X86II::MRM5m: + case X86II::MRM6m: case X86II::MRM7m: + case X86II::MRMSrcMem: + case X86II::MRMSrcReg: + if (MI.getNumOperands() > CurOp && MI.getOperand(CurOp).isReg() && + X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_R = 0x0; + CurOp++; + + if (HasVEX_4V) { + VEX_4V = getVEXRegisterEncoding(MI, IsDestMem ? CurOp-1 : CurOp); + CurOp++; + } + + // To only check operands before the memory address ones, start + // the search from the beginning + if (IsDestMem) + CurOp = 0; + + // If the last register should be encoded in the immediate field + // do not use any bit from VEX prefix to this register, ignore it + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) + NumOps--; + + for (; CurOp != NumOps; ++CurOp) { + const MCOperand &MO = MI.getOperand(CurOp); + if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg())) + VEX_B = 0x0; + if (!VEX_B && MO.isReg() && + ((TSFlags & X86II::FormMask) == X86II::MRMSrcMem) && + X86II::isX86_64ExtendedReg(MO.getReg())) + VEX_X = 0x0; + } + break; + default: // MRMDestReg, MRM0r-MRM7r, RawFrm + if (!MI.getNumOperands()) + break; + + if (MI.getOperand(CurOp).isReg() && + X86II::isX86_64ExtendedReg(MI.getOperand(CurOp).getReg())) + VEX_B = 0; + + if (HasVEX_4V) + VEX_4V = getVEXRegisterEncoding(MI, CurOp); + + CurOp++; + for (; CurOp != NumOps; ++CurOp) { + const MCOperand &MO = MI.getOperand(CurOp); + if (MO.isReg() && !HasVEX_4V && + X86II::isX86_64ExtendedReg(MO.getReg())) + VEX_R = 0x0; + } + break; + } + + // Emit segment override opcode prefix as needed. + EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS); + + // VEX opcode prefix can have 2 or 3 bytes + // + // 3 bytes: + // +-----+ +--------------+ +-------------------+ + // | C4h | | RXB | m-mmmm | | W | vvvv | L | pp | + // +-----+ +--------------+ +-------------------+ + // 2 bytes: + // +-----+ +-------------------+ + // | C5h | | R | vvvv | L | pp | + // +-----+ +-------------------+ + // + unsigned char LastByte = VEX_PP | (VEX_L << 2) | (VEX_4V << 3); + + if (VEX_B && VEX_X && !VEX_W && (VEX_5M == 1)) { // 2 byte VEX prefix + EmitByte(0xC5, CurByte, OS); + EmitByte(LastByte | (VEX_R << 7), CurByte, OS); + return; + } + + // 3 byte VEX prefix + EmitByte(0xC4, CurByte, OS); + EmitByte(VEX_R << 7 | VEX_X << 6 | VEX_B << 5 | VEX_5M, CurByte, OS); + EmitByte(LastByte | (VEX_W << 7), CurByte, OS); +} + +/// DetermineREXPrefix - Determine if the MCInst has to be encoded with a X86-64 +/// REX prefix which specifies 1) 64-bit instructions, 2) non-default operand +/// size, and 3) use of X86-64 extended registers. +static unsigned DetermineREXPrefix(const MCInst &MI, uint64_t TSFlags, + const MCInstrDesc &Desc) { + unsigned REX = 0; + if (TSFlags & X86II::REX_W) + REX |= 1 << 3; // set REX.W + + if (MI.getNumOperands() == 0) return REX; + + unsigned NumOps = MI.getNumOperands(); + // FIXME: MCInst should explicitize the two-addrness. + bool isTwoAddr = NumOps > 1 && + Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1; + + // If it accesses SPL, BPL, SIL, or DIL, then it requires a 0x40 REX prefix. + unsigned i = isTwoAddr ? 1 : 0; + for (; i != NumOps; ++i) { + const MCOperand &MO = MI.getOperand(i); + if (!MO.isReg()) continue; + unsigned Reg = MO.getReg(); + if (!X86II::isX86_64NonExtLowByteReg(Reg)) continue; + // FIXME: The caller of DetermineREXPrefix slaps this prefix onto anything + // that returns non-zero. + REX |= 0x40; // REX fixed encoding prefix + break; + } + + switch (TSFlags & X86II::FormMask) { + case X86II::MRMInitReg: assert(0 && "FIXME: Remove this!"); + case X86II::MRMSrcReg: + if (MI.getOperand(0).isReg() && + X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg())) + REX |= 1 << 2; // set REX.R + i = isTwoAddr ? 2 : 1; + for (; i != NumOps; ++i) { + const MCOperand &MO = MI.getOperand(i); + if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg())) + REX |= 1 << 0; // set REX.B + } + break; + case X86II::MRMSrcMem: { + if (MI.getOperand(0).isReg() && + X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg())) + REX |= 1 << 2; // set REX.R + unsigned Bit = 0; + i = isTwoAddr ? 2 : 1; + for (; i != NumOps; ++i) { + const MCOperand &MO = MI.getOperand(i); + if (MO.isReg()) { + if (X86II::isX86_64ExtendedReg(MO.getReg())) + REX |= 1 << Bit; // set REX.B (Bit=0) and REX.X (Bit=1) + Bit++; + } + } + break; + } + case X86II::MRM0m: case X86II::MRM1m: + case X86II::MRM2m: case X86II::MRM3m: + case X86II::MRM4m: case X86II::MRM5m: + case X86II::MRM6m: case X86II::MRM7m: + case X86II::MRMDestMem: { + unsigned e = (isTwoAddr ? X86::AddrNumOperands+1 : X86::AddrNumOperands); + i = isTwoAddr ? 1 : 0; + if (NumOps > e && MI.getOperand(e).isReg() && + X86II::isX86_64ExtendedReg(MI.getOperand(e).getReg())) + REX |= 1 << 2; // set REX.R + unsigned Bit = 0; + for (; i != e; ++i) { + const MCOperand &MO = MI.getOperand(i); + if (MO.isReg()) { + if (X86II::isX86_64ExtendedReg(MO.getReg())) + REX |= 1 << Bit; // REX.B (Bit=0) and REX.X (Bit=1) + Bit++; + } + } + break; + } + default: + if (MI.getOperand(0).isReg() && + X86II::isX86_64ExtendedReg(MI.getOperand(0).getReg())) + REX |= 1 << 0; // set REX.B + i = isTwoAddr ? 2 : 1; + for (unsigned e = NumOps; i != e; ++i) { + const MCOperand &MO = MI.getOperand(i); + if (MO.isReg() && X86II::isX86_64ExtendedReg(MO.getReg())) + REX |= 1 << 2; // set REX.R + } + break; + } + return REX; +} + +/// EmitSegmentOverridePrefix - Emit segment override opcode prefix as needed +void X86MCCodeEmitter::EmitSegmentOverridePrefix(uint64_t TSFlags, + unsigned &CurByte, int MemOperand, + const MCInst &MI, + raw_ostream &OS) const { + switch (TSFlags & X86II::SegOvrMask) { + default: assert(0 && "Invalid segment!"); + case 0: + // No segment override, check for explicit one on memory operand. + if (MemOperand != -1) { // If the instruction has a memory operand. + switch (MI.getOperand(MemOperand+X86::AddrSegmentReg).getReg()) { + default: assert(0 && "Unknown segment register!"); + case 0: break; + case X86::CS: EmitByte(0x2E, CurByte, OS); break; + case X86::SS: EmitByte(0x36, CurByte, OS); break; + case X86::DS: EmitByte(0x3E, CurByte, OS); break; + case X86::ES: EmitByte(0x26, CurByte, OS); break; + case X86::FS: EmitByte(0x64, CurByte, OS); break; + case X86::GS: EmitByte(0x65, CurByte, OS); break; + } + } + break; + case X86II::FS: + EmitByte(0x64, CurByte, OS); + break; + case X86II::GS: + EmitByte(0x65, CurByte, OS); + break; + } +} + +/// EmitOpcodePrefix - Emit all instruction prefixes prior to the opcode. +/// +/// MemOperand is the operand # of the start of a memory operand if present. If +/// Not present, it is -1. +void X86MCCodeEmitter::EmitOpcodePrefix(uint64_t TSFlags, unsigned &CurByte, + int MemOperand, const MCInst &MI, + const MCInstrDesc &Desc, + raw_ostream &OS) const { + + // Emit the lock opcode prefix as needed. + if (TSFlags & X86II::LOCK) + EmitByte(0xF0, CurByte, OS); + + // Emit segment override opcode prefix as needed. + EmitSegmentOverridePrefix(TSFlags, CurByte, MemOperand, MI, OS); + + // Emit the repeat opcode prefix as needed. + if ((TSFlags & X86II::Op0Mask) == X86II::REP) + EmitByte(0xF3, CurByte, OS); + + // Emit the address size opcode prefix as needed. + if ((TSFlags & X86II::AdSize) || + (MemOperand != -1 && is64BitMode() && Is32BitMemOperand(MI, MemOperand))) + EmitByte(0x67, CurByte, OS); + + // Emit the operand size opcode prefix as needed. + if (TSFlags & X86II::OpSize) + EmitByte(0x66, CurByte, OS); + + bool Need0FPrefix = false; + switch (TSFlags & X86II::Op0Mask) { + default: assert(0 && "Invalid prefix!"); + case 0: break; // No prefix! + case X86II::REP: break; // already handled. + case X86II::TB: // Two-byte opcode prefix + case X86II::T8: // 0F 38 + case X86II::TA: // 0F 3A + case X86II::A6: // 0F A6 + case X86II::A7: // 0F A7 + Need0FPrefix = true; + break; + case X86II::TF: // F2 0F 38 + EmitByte(0xF2, CurByte, OS); + Need0FPrefix = true; + break; + case X86II::XS: // F3 0F + EmitByte(0xF3, CurByte, OS); + Need0FPrefix = true; + break; + case X86II::XD: // F2 0F + EmitByte(0xF2, CurByte, OS); + Need0FPrefix = true; + break; + case X86II::D8: EmitByte(0xD8, CurByte, OS); break; + case X86II::D9: EmitByte(0xD9, CurByte, OS); break; + case X86II::DA: EmitByte(0xDA, CurByte, OS); break; + case X86II::DB: EmitByte(0xDB, CurByte, OS); break; + case X86II::DC: EmitByte(0xDC, CurByte, OS); break; + case X86II::DD: EmitByte(0xDD, CurByte, OS); break; + case X86II::DE: EmitByte(0xDE, CurByte, OS); break; + case X86II::DF: EmitByte(0xDF, CurByte, OS); break; + } + + // Handle REX prefix. + // FIXME: Can this come before F2 etc to simplify emission? + if (is64BitMode()) { + if (unsigned REX = DetermineREXPrefix(MI, TSFlags, Desc)) + EmitByte(0x40 | REX, CurByte, OS); + } + + // 0x0F escape code must be emitted just before the opcode. + if (Need0FPrefix) + EmitByte(0x0F, CurByte, OS); + + // FIXME: Pull this up into previous switch if REX can be moved earlier. + switch (TSFlags & X86II::Op0Mask) { + case X86II::TF: // F2 0F 38 + case X86II::T8: // 0F 38 + EmitByte(0x38, CurByte, OS); + break; + case X86II::TA: // 0F 3A + EmitByte(0x3A, CurByte, OS); + break; + case X86II::A6: // 0F A6 + EmitByte(0xA6, CurByte, OS); + break; + case X86II::A7: // 0F A7 + EmitByte(0xA7, CurByte, OS); + break; + } +} + +void X86MCCodeEmitter:: +EncodeInstruction(const MCInst &MI, raw_ostream &OS, + SmallVectorImpl<MCFixup> &Fixups) const { + unsigned Opcode = MI.getOpcode(); + const MCInstrDesc &Desc = MCII.get(Opcode); + uint64_t TSFlags = Desc.TSFlags; + + // Pseudo instructions don't get encoded. + if ((TSFlags & X86II::FormMask) == X86II::Pseudo) + return; + + // If this is a two-address instruction, skip one of the register operands. + // FIXME: This should be handled during MCInst lowering. + unsigned NumOps = Desc.getNumOperands(); + unsigned CurOp = 0; + if (NumOps > 1 && Desc.getOperandConstraint(1, MCOI::TIED_TO) != -1) + ++CurOp; + else if (NumOps > 2 && Desc.getOperandConstraint(NumOps-1, MCOI::TIED_TO)== 0) + // Skip the last source operand that is tied_to the dest reg. e.g. LXADD32 + --NumOps; + + // Keep track of the current byte being emitted. + unsigned CurByte = 0; + + // Is this instruction encoded using the AVX VEX prefix? + bool HasVEXPrefix = false; + + // It uses the VEX.VVVV field? + bool HasVEX_4V = false; + + if ((TSFlags >> X86II::VEXShift) & X86II::VEX) + HasVEXPrefix = true; + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_4V) + HasVEX_4V = true; + + + // Determine where the memory operand starts, if present. + int MemoryOperand = X86II::getMemoryOperandNo(TSFlags); + if (MemoryOperand != -1) MemoryOperand += CurOp; + + if (!HasVEXPrefix) + EmitOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS); + else + EmitVEXOpcodePrefix(TSFlags, CurByte, MemoryOperand, MI, Desc, OS); + + + unsigned char BaseOpcode = X86II::getBaseOpcodeFor(TSFlags); + + if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode) + BaseOpcode = 0x0F; // Weird 3DNow! encoding. + + unsigned SrcRegNum = 0; + switch (TSFlags & X86II::FormMask) { + case X86II::MRMInitReg: + assert(0 && "FIXME: Remove this form when the JIT moves to MCCodeEmitter!"); + default: errs() << "FORM: " << (TSFlags & X86II::FormMask) << "\n"; + assert(0 && "Unknown FormMask value in X86MCCodeEmitter!"); + case X86II::Pseudo: + assert(0 && "Pseudo instruction shouldn't be emitted"); + case X86II::RawFrm: + EmitByte(BaseOpcode, CurByte, OS); + break; + + case X86II::RawFrmImm8: + EmitByte(BaseOpcode, CurByte, OS); + EmitImmediate(MI.getOperand(CurOp++), + X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags), + CurByte, OS, Fixups); + EmitImmediate(MI.getOperand(CurOp++), 1, FK_Data_1, CurByte, OS, Fixups); + break; + case X86II::RawFrmImm16: + EmitByte(BaseOpcode, CurByte, OS); + EmitImmediate(MI.getOperand(CurOp++), + X86II::getSizeOfImm(TSFlags), getImmFixupKind(TSFlags), + CurByte, OS, Fixups); + EmitImmediate(MI.getOperand(CurOp++), 2, FK_Data_2, CurByte, OS, Fixups); + break; + + case X86II::AddRegFrm: + EmitByte(BaseOpcode + GetX86RegNum(MI.getOperand(CurOp++)), CurByte, OS); + break; + + case X86II::MRMDestReg: + EmitByte(BaseOpcode, CurByte, OS); + EmitRegModRMByte(MI.getOperand(CurOp), + GetX86RegNum(MI.getOperand(CurOp+1)), CurByte, OS); + CurOp += 2; + break; + + case X86II::MRMDestMem: + EmitByte(BaseOpcode, CurByte, OS); + SrcRegNum = CurOp + X86::AddrNumOperands; + + if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV) + SrcRegNum++; + + EmitMemModRMByte(MI, CurOp, + GetX86RegNum(MI.getOperand(SrcRegNum)), + TSFlags, CurByte, OS, Fixups); + CurOp = SrcRegNum + 1; + break; + + case X86II::MRMSrcReg: + EmitByte(BaseOpcode, CurByte, OS); + SrcRegNum = CurOp + 1; + + if (HasVEX_4V) // Skip 1st src (which is encoded in VEX_VVVV) + SrcRegNum++; + + EmitRegModRMByte(MI.getOperand(SrcRegNum), + GetX86RegNum(MI.getOperand(CurOp)), CurByte, OS); + CurOp = SrcRegNum + 1; + break; + + case X86II::MRMSrcMem: { + int AddrOperands = X86::AddrNumOperands; + unsigned FirstMemOp = CurOp+1; + if (HasVEX_4V) { + ++AddrOperands; + ++FirstMemOp; // Skip the register source (which is encoded in VEX_VVVV). + } + + EmitByte(BaseOpcode, CurByte, OS); + + EmitMemModRMByte(MI, FirstMemOp, GetX86RegNum(MI.getOperand(CurOp)), + TSFlags, CurByte, OS, Fixups); + CurOp += AddrOperands + 1; + break; + } + + case X86II::MRM0r: case X86II::MRM1r: + case X86II::MRM2r: case X86II::MRM3r: + case X86II::MRM4r: case X86II::MRM5r: + case X86II::MRM6r: case X86II::MRM7r: + if (HasVEX_4V) // Skip the register dst (which is encoded in VEX_VVVV). + CurOp++; + EmitByte(BaseOpcode, CurByte, OS); + EmitRegModRMByte(MI.getOperand(CurOp++), + (TSFlags & X86II::FormMask)-X86II::MRM0r, + CurByte, OS); + break; + case X86II::MRM0m: case X86II::MRM1m: + case X86II::MRM2m: case X86II::MRM3m: + case X86II::MRM4m: case X86II::MRM5m: + case X86II::MRM6m: case X86II::MRM7m: + EmitByte(BaseOpcode, CurByte, OS); + EmitMemModRMByte(MI, CurOp, (TSFlags & X86II::FormMask)-X86II::MRM0m, + TSFlags, CurByte, OS, Fixups); + CurOp += X86::AddrNumOperands; + break; + case X86II::MRM_C1: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xC1, CurByte, OS); + break; + case X86II::MRM_C2: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xC2, CurByte, OS); + break; + case X86II::MRM_C3: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xC3, CurByte, OS); + break; + case X86II::MRM_C4: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xC4, CurByte, OS); + break; + case X86II::MRM_C8: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xC8, CurByte, OS); + break; + case X86II::MRM_C9: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xC9, CurByte, OS); + break; + case X86II::MRM_E8: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xE8, CurByte, OS); + break; + case X86II::MRM_F0: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xF0, CurByte, OS); + break; + case X86II::MRM_F8: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xF8, CurByte, OS); + break; + case X86II::MRM_F9: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xF9, CurByte, OS); + break; + case X86II::MRM_D0: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xD0, CurByte, OS); + break; + case X86II::MRM_D1: + EmitByte(BaseOpcode, CurByte, OS); + EmitByte(0xD1, CurByte, OS); + break; + } + + // If there is a remaining operand, it must be a trailing immediate. Emit it + // according to the right size for the instruction. + if (CurOp != NumOps) { + // The last source register of a 4 operand instruction in AVX is encoded + // in bits[7:4] of a immediate byte, and bits[3:0] are ignored. + if ((TSFlags >> X86II::VEXShift) & X86II::VEX_I8IMM) { + const MCOperand &MO = MI.getOperand(CurOp++); + bool IsExtReg = + X86II::isX86_64ExtendedReg(MO.getReg()); + unsigned RegNum = (IsExtReg ? (1 << 7) : 0); + RegNum |= GetX86RegNum(MO) << 4; + EmitImmediate(MCOperand::CreateImm(RegNum), 1, FK_Data_1, CurByte, OS, + Fixups); + } else { + unsigned FixupKind; + // FIXME: Is there a better way to know that we need a signed relocation? + if (MI.getOpcode() == X86::ADD64ri32 || + MI.getOpcode() == X86::MOV64ri32 || + MI.getOpcode() == X86::MOV64mi32 || + MI.getOpcode() == X86::PUSH64i32) + FixupKind = X86::reloc_signed_4byte; + else + FixupKind = getImmFixupKind(TSFlags); + EmitImmediate(MI.getOperand(CurOp++), + X86II::getSizeOfImm(TSFlags), MCFixupKind(FixupKind), + CurByte, OS, Fixups); + } + } + + if ((TSFlags >> X86II::VEXShift) & X86II::Has3DNow0F0FOpcode) + EmitByte(X86II::getBaseOpcodeFor(TSFlags), CurByte, OS); + + +#ifndef NDEBUG + // FIXME: Verify. + if (/*!Desc.isVariadic() &&*/ CurOp != NumOps) { + errs() << "Cannot encode all operands of: "; + MI.dump(); + errs() << '\n'; + abort(); + } +#endif +} diff --git a/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.cpp b/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.cpp index 75bdc508fa8..129c3c91b3c 100644 --- a/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.cpp +++ b/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.cpp @@ -16,6 +16,7 @@ #include "llvm/MC/MachineLocation.h" #include "llvm/MC/MCInstrInfo.h" #include "llvm/MC/MCRegisterInfo.h" +#include "llvm/MC/MCStreamer.h" #include "llvm/MC/MCSubtargetInfo.h" #include "llvm/Target/TargetRegistry.h" #include "llvm/ADT/Triple.h" @@ -363,6 +364,23 @@ static MCCodeGenInfo *createX86MCCodeGenInfo(StringRef TT, Reloc::Model RM, return X; } +static MCStreamer *createMCStreamer(const Target &T, const std::string &TT, + MCContext &Ctx, TargetAsmBackend &TAB, + raw_ostream &_OS, + MCCodeEmitter *_Emitter, + bool RelaxAll, + bool NoExecStack) { + Triple TheTriple(TT); + + if (TheTriple.isOSDarwin() || TheTriple.getEnvironment() == Triple::MachO) + return createMachOStreamer(Ctx, TAB, _OS, _Emitter, RelaxAll); + + if (TheTriple.isOSWindows()) + return createWinCOFFStreamer(Ctx, TAB, *_Emitter, _OS, RelaxAll); + + return createELFStreamer(Ctx, TAB, _OS, _Emitter, RelaxAll, NoExecStack); +} + // Force static initialization. extern "C" void LLVMInitializeX86TargetMC() { // Register the MC asm info. @@ -386,4 +404,22 @@ extern "C" void LLVMInitializeX86TargetMC() { X86_MC::createX86MCSubtargetInfo); TargetRegistry::RegisterMCSubtargetInfo(TheX86_64Target, X86_MC::createX86MCSubtargetInfo); + + // Register the code emitter. + TargetRegistry::RegisterCodeEmitter(TheX86_32Target, + createX86MCCodeEmitter); + TargetRegistry::RegisterCodeEmitter(TheX86_64Target, + createX86MCCodeEmitter); + + // Register the asm backend. + TargetRegistry::RegisterAsmBackend(TheX86_32Target, + createX86_32AsmBackend); + TargetRegistry::RegisterAsmBackend(TheX86_64Target, + createX86_64AsmBackend); + + // Register the object streamer. + TargetRegistry::RegisterObjectStreamer(TheX86_32Target, + createMCStreamer); + TargetRegistry::RegisterObjectStreamer(TheX86_64Target, + createMCStreamer); } diff --git a/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.h b/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.h index 29b9cd9b771..dd893be7ec8 100644 --- a/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.h +++ b/llvm/lib/Target/X86/MCTargetDesc/X86MCTargetDesc.h @@ -20,10 +20,13 @@ namespace llvm { class MCCodeEmitter; class MCContext; class MCInstrInfo; +class MCObjectWriter; class MCRegisterInfo; class MCSubtargetInfo; class Target; +class TargetAsmBackend; class StringRef; +class raw_ostream; extern Target TheX86_32Target, TheX86_64Target; @@ -70,6 +73,14 @@ MCCodeEmitter *createX86MCCodeEmitter(const MCInstrInfo &MCII, const MCSubtargetInfo &STI, MCContext &Ctx); +TargetAsmBackend *createX86_32AsmBackend(const Target &, const std::string &); +TargetAsmBackend *createX86_64AsmBackend(const Target &, const std::string &); + +/// createX86MachObjectWriter - Construct an X86 Mach-O object writer. +MCObjectWriter *createX86MachObjectWriter(raw_ostream &OS, + bool Is64Bit, + uint32_t CPUType, + uint32_t CPUSubtype); } // End llvm namespace diff --git a/llvm/lib/Target/X86/MCTargetDesc/X86MachObjectWriter.cpp b/llvm/lib/Target/X86/MCTargetDesc/X86MachObjectWriter.cpp new file mode 100644 index 00000000000..f0f1982d57f --- /dev/null +++ b/llvm/lib/Target/X86/MCTargetDesc/X86MachObjectWriter.cpp @@ -0,0 +1,554 @@ +//===-- X86MachObjectWriter.cpp - X86 Mach-O Writer -----------------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// + +#include "MCTargetDesc/X86FixupKinds.h" +#include "MCTargetDesc/X86MCTargetDesc.h" +#include "llvm/MC/MCAssembler.h" +#include "llvm/MC/MCAsmLayout.h" +#include "llvm/MC/MCMachObjectWriter.h" +#include "llvm/MC/MCSectionMachO.h" +#include "llvm/MC/MCValue.h" +#include "llvm/ADT/Twine.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Object/MachOFormat.h" + +using namespace llvm; +using namespace llvm::object; + +namespace { +class X86MachObjectWriter : public MCMachObjectTargetWriter { + void RecordScatteredRelocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + unsigned Log2Size, + uint64_t &FixedValue); + void RecordTLVPRelocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + uint64_t &FixedValue); + + void RecordX86Relocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + uint64_t &FixedValue); + void RecordX86_64Relocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + uint64_t &FixedValue); +public: + X86MachObjectWriter(bool Is64Bit, uint32_t CPUType, + uint32_t CPUSubtype) + : MCMachObjectTargetWriter(Is64Bit, CPUType, CPUSubtype, + /*UseAggressiveSymbolFolding=*/Is64Bit) {} + + void RecordRelocation(MachObjectWriter *Writer, + const MCAssembler &Asm, const MCAsmLayout &Layout, + const MCFragment *Fragment, const MCFixup &Fixup, + MCValue Target, uint64_t &FixedValue) { + if (Writer->is64Bit()) + RecordX86_64Relocation(Writer, Asm, Layout, Fragment, Fixup, Target, + FixedValue); + else + RecordX86Relocation(Writer, Asm, Layout, Fragment, Fixup, Target, + FixedValue); + } +}; +} + +static bool isFixupKindRIPRel(unsigned Kind) { + return Kind == X86::reloc_riprel_4byte || + Kind == X86::reloc_riprel_4byte_movq_load; +} + +static unsigned getFixupKindLog2Size(unsigned Kind) { + switch (Kind) { + default: + llvm_unreachable("invalid fixup kind!"); + case FK_PCRel_1: + case FK_Data_1: return 0; + case FK_PCRel_2: + case FK_Data_2: return 1; + case FK_PCRel_4: + // FIXME: Remove these!!! + case X86::reloc_riprel_4byte: + case X86::reloc_riprel_4byte_movq_load: + case X86::reloc_signed_4byte: + case FK_Data_4: return 2; + case FK_Data_8: return 3; + } +} + +void X86MachObjectWriter::RecordX86_64Relocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + uint64_t &FixedValue) { + unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind()); + unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind()); + unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind()); + + // See <reloc.h>. + uint32_t FixupOffset = + Layout.getFragmentOffset(Fragment) + Fixup.getOffset(); + uint32_t FixupAddress = + Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset(); + int64_t Value = 0; + unsigned Index = 0; + unsigned IsExtern = 0; + unsigned Type = 0; + + Value = Target.getConstant(); + + if (IsPCRel) { + // Compensate for the relocation offset, Darwin x86_64 relocations only have + // the addend and appear to have attempted to define it to be the actual + // expression addend without the PCrel bias. However, instructions with data + // following the relocation are not accommodated for (see comment below + // regarding SIGNED{1,2,4}), so it isn't exactly that either. + Value += 1LL << Log2Size; + } + + if (Target.isAbsolute()) { // constant + // SymbolNum of 0 indicates the absolute section. + Type = macho::RIT_X86_64_Unsigned; + Index = 0; + + // FIXME: I believe this is broken, I don't think the linker can understand + // it. I think it would require a local relocation, but I'm not sure if that + // would work either. The official way to get an absolute PCrel relocation + // is to use an absolute symbol (which we don't support yet). + if (IsPCRel) { + IsExtern = 1; + Type = macho::RIT_X86_64_Branch; + } + } else if (Target.getSymB()) { // A - B + constant + const MCSymbol *A = &Target.getSymA()->getSymbol(); + MCSymbolData &A_SD = Asm.getSymbolData(*A); + const MCSymbolData *A_Base = Asm.getAtom(&A_SD); + + const MCSymbol *B = &Target.getSymB()->getSymbol(); + MCSymbolData &B_SD = Asm.getSymbolData(*B); + const MCSymbolData *B_Base = Asm.getAtom(&B_SD); + + // Neither symbol can be modified. + if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None || + Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None) + report_fatal_error("unsupported relocation of modified symbol"); + + // We don't support PCrel relocations of differences. Darwin 'as' doesn't + // implement most of these correctly. + if (IsPCRel) + report_fatal_error("unsupported pc-relative relocation of difference"); + + // The support for the situation where one or both of the symbols would + // require a local relocation is handled just like if the symbols were + // external. This is certainly used in the case of debug sections where the + // section has only temporary symbols and thus the symbols don't have base + // symbols. This is encoded using the section ordinal and non-extern + // relocation entries. + + // Darwin 'as' doesn't emit correct relocations for this (it ends up with a + // single SIGNED relocation); reject it for now. Except the case where both + // symbols don't have a base, equal but both NULL. + if (A_Base == B_Base && A_Base) + report_fatal_error("unsupported relocation with identical base"); + + Value += Writer->getSymbolAddress(&A_SD, Layout) - + (A_Base == NULL ? 0 : Writer->getSymbolAddress(A_Base, Layout)); + Value -= Writer->getSymbolAddress(&B_SD, Layout) - + (B_Base == NULL ? 0 : Writer->getSymbolAddress(B_Base, Layout)); + + if (A_Base) { + Index = A_Base->getIndex(); + IsExtern = 1; + } + else { + Index = A_SD.getFragment()->getParent()->getOrdinal() + 1; + IsExtern = 0; + } + Type = macho::RIT_X86_64_Unsigned; + + macho::RelocationEntry MRE; + MRE.Word0 = FixupOffset; + MRE.Word1 = ((Index << 0) | + (IsPCRel << 24) | + (Log2Size << 25) | + (IsExtern << 27) | + (Type << 28)); + Writer->addRelocation(Fragment->getParent(), MRE); + + if (B_Base) { + Index = B_Base->getIndex(); + IsExtern = 1; + } + else { + Index = B_SD.getFragment()->getParent()->getOrdinal() + 1; + IsExtern = 0; + } + Type = macho::RIT_X86_64_Subtractor; + } else { + const MCSymbol *Symbol = &Target.getSymA()->getSymbol(); + MCSymbolData &SD = Asm.getSymbolData(*Symbol); + const MCSymbolData *Base = Asm.getAtom(&SD); + + // Relocations inside debug sections always use local relocations when + // possible. This seems to be done because the debugger doesn't fully + // understand x86_64 relocation entries, and expects to find values that + // have already been fixed up. + if (Symbol->isInSection()) { + const MCSectionMachO &Section = static_cast<const MCSectionMachO&>( + Fragment->getParent()->getSection()); + if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG)) + Base = 0; + } + + // x86_64 almost always uses external relocations, except when there is no + // symbol to use as a base address (a local symbol with no preceding + // non-local symbol). + if (Base) { + Index = Base->getIndex(); + IsExtern = 1; + + // Add the local offset, if needed. + if (Base != &SD) + Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base); + } else if (Symbol->isInSection() && !Symbol->isVariable()) { + // The index is the section ordinal (1-based). + Index = SD.getFragment()->getParent()->getOrdinal() + 1; + IsExtern = 0; + Value += Writer->getSymbolAddress(&SD, Layout); + + if (IsPCRel) + Value -= FixupAddress + (1 << Log2Size); + } else if (Symbol->isVariable()) { + const MCExpr *Value = Symbol->getVariableValue(); + int64_t Res; + bool isAbs = Value->EvaluateAsAbsolute(Res, Layout, + Writer->getSectionAddressMap()); + if (isAbs) { + FixedValue = Res; + return; + } else { + report_fatal_error("unsupported relocation of variable '" + + Symbol->getName() + "'"); + } + } else { + report_fatal_error("unsupported relocation of undefined symbol '" + + Symbol->getName() + "'"); + } + + MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind(); + if (IsPCRel) { + if (IsRIPRel) { + if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) { + // x86_64 distinguishes movq foo@GOTPCREL so that the linker can + // rewrite the movq to an leaq at link time if the symbol ends up in + // the same linkage unit. + if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load) + Type = macho::RIT_X86_64_GOTLoad; + else + Type = macho::RIT_X86_64_GOT; + } else if (Modifier == MCSymbolRefExpr::VK_TLVP) { + Type = macho::RIT_X86_64_TLV; + } else if (Modifier != MCSymbolRefExpr::VK_None) { + report_fatal_error("unsupported symbol modifier in relocation"); + } else { + Type = macho::RIT_X86_64_Signed; + + // The Darwin x86_64 relocation format has a problem where it cannot + // encode an address (L<foo> + <constant>) which is outside the atom + // containing L<foo>. Generally, this shouldn't occur but it does + // happen when we have a RIPrel instruction with data following the + // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel + // adjustment Darwin x86_64 uses, the offset is still negative and the + // linker has no way to recognize this. + // + // To work around this, Darwin uses several special relocation types + // to indicate the offsets. However, the specification or + // implementation of these seems to also be incomplete; they should + // adjust the addend as well based on the actual encoded instruction + // (the additional bias), but instead appear to just look at the final + // offset. + switch (-(Target.getConstant() + (1LL << Log2Size))) { + case 1: Type = macho::RIT_X86_64_Signed1; break; + case 2: Type = macho::RIT_X86_64_Signed2; break; + case 4: Type = macho::RIT_X86_64_Signed4; break; + } + } + } else { + if (Modifier != MCSymbolRefExpr::VK_None) + report_fatal_error("unsupported symbol modifier in branch " + "relocation"); + + Type = macho::RIT_X86_64_Branch; + } + } else { + if (Modifier == MCSymbolRefExpr::VK_GOT) { + Type = macho::RIT_X86_64_GOT; + } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) { + // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which + // case all we do is set the PCrel bit in the relocation entry; this is + // used with exception handling, for example. The source is required to + // include any necessary offset directly. + Type = macho::RIT_X86_64_GOT; + IsPCRel = 1; + } else if (Modifier == MCSymbolRefExpr::VK_TLVP) { + report_fatal_error("TLVP symbol modifier should have been rip-rel"); + } else if (Modifier != MCSymbolRefExpr::VK_None) + report_fatal_error("unsupported symbol modifier in relocation"); + else + Type = macho::RIT_X86_64_Unsigned; + } + } + + // x86_64 always writes custom values into the fixups. + FixedValue = Value; + + // struct relocation_info (8 bytes) + macho::RelocationEntry MRE; + MRE.Word0 = FixupOffset; + MRE.Word1 = ((Index << 0) | + (IsPCRel << 24) | + (Log2Size << 25) | + (IsExtern << 27) | + (Type << 28)); + Writer->addRelocation(Fragment->getParent(), MRE); +} + +void X86MachObjectWriter::RecordScatteredRelocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + unsigned Log2Size, + uint64_t &FixedValue) { + uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset(); + unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind()); + unsigned Type = macho::RIT_Vanilla; + + // See <reloc.h>. + const MCSymbol *A = &Target.getSymA()->getSymbol(); + MCSymbolData *A_SD = &Asm.getSymbolData(*A); + + if (!A_SD->getFragment()) + report_fatal_error("symbol '" + A->getName() + + "' can not be undefined in a subtraction expression"); + + uint32_t Value = Writer->getSymbolAddress(A_SD, Layout); + uint64_t SecAddr = Writer->getSectionAddress(A_SD->getFragment()->getParent()); + FixedValue += SecAddr; + uint32_t Value2 = 0; + + if (const MCSymbolRefExpr *B = Target.getSymB()) { + MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol()); + + if (!B_SD->getFragment()) + report_fatal_error("symbol '" + B->getSymbol().getName() + + "' can not be undefined in a subtraction expression"); + + // Select the appropriate difference relocation type. + // + // Note that there is no longer any semantic difference between these two + // relocation types from the linkers point of view, this is done solely for + // pedantic compatibility with 'as'. + Type = A_SD->isExternal() ? (unsigned)macho::RIT_Difference : + (unsigned)macho::RIT_Generic_LocalDifference; + Value2 = Writer->getSymbolAddress(B_SD, Layout); + FixedValue -= Writer->getSectionAddress(B_SD->getFragment()->getParent()); + } + + // Relocations are written out in reverse order, so the PAIR comes first. + if (Type == macho::RIT_Difference || + Type == macho::RIT_Generic_LocalDifference) { + macho::RelocationEntry MRE; + MRE.Word0 = ((0 << 0) | + (macho::RIT_Pair << 24) | + (Log2Size << 28) | + (IsPCRel << 30) | + macho::RF_Scattered); + MRE.Word1 = Value2; + Writer->addRelocation(Fragment->getParent(), MRE); + } + + macho::RelocationEntry MRE; + MRE.Word0 = ((FixupOffset << 0) | + (Type << 24) | + (Log2Size << 28) | + (IsPCRel << 30) | + macho::RF_Scattered); + MRE.Word1 = Value; + Writer->addRelocation(Fragment->getParent(), MRE); +} + +void X86MachObjectWriter::RecordTLVPRelocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + uint64_t &FixedValue) { + assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP && + !is64Bit() && + "Should only be called with a 32-bit TLVP relocation!"); + + unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind()); + uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset(); + unsigned IsPCRel = 0; + + // Get the symbol data. + MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol()); + unsigned Index = SD_A->getIndex(); + + // We're only going to have a second symbol in pic mode and it'll be a + // subtraction from the picbase. For 32-bit pic the addend is the difference + // between the picbase and the next address. For 32-bit static the addend is + // zero. + if (Target.getSymB()) { + // If this is a subtraction then we're pcrel. + uint32_t FixupAddress = + Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset(); + MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol()); + IsPCRel = 1; + FixedValue = (FixupAddress - Writer->getSymbolAddress(SD_B, Layout) + + Target.getConstant()); + FixedValue += 1ULL << Log2Size; + } else { + FixedValue = 0; + } + + // struct relocation_info (8 bytes) + macho::RelocationEntry MRE; + MRE.Word0 = Value; + MRE.Word1 = ((Index << 0) | + (IsPCRel << 24) | + (Log2Size << 25) | + (1 << 27) | // Extern + (macho::RIT_Generic_TLV << 28)); // Type + Writer->addRelocation(Fragment->getParent(), MRE); +} + +void X86MachObjectWriter::RecordX86Relocation(MachObjectWriter *Writer, + const MCAssembler &Asm, + const MCAsmLayout &Layout, + const MCFragment *Fragment, + const MCFixup &Fixup, + MCValue Target, + uint64_t &FixedValue) { + unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind()); + unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind()); + + // If this is a 32-bit TLVP reloc it's handled a bit differently. + if (Target.getSymA() && + Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) { + RecordTLVPRelocation(Writer, Asm, Layout, Fragment, Fixup, Target, + FixedValue); + return; + } + + // If this is a difference or a defined symbol plus an offset, then we need a + // scattered relocation entry. Differences always require scattered + // relocations. + if (Target.getSymB()) + return RecordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup, + Target, Log2Size, FixedValue); + + // Get the symbol data, if any. + MCSymbolData *SD = 0; + if (Target.getSymA()) + SD = &Asm.getSymbolData(Target.getSymA()->getSymbol()); + + // If this is an internal relocation with an offset, it also needs a scattered + // relocation entry. + uint32_t Offset = Target.getConstant(); + if (IsPCRel) + Offset += 1 << Log2Size; + if (Offset && SD && !Writer->doesSymbolRequireExternRelocation(SD)) + return RecordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup, + Target, Log2Size, FixedValue); + + // See <reloc.h>. + uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset(); + unsigned Index = 0; + unsigned IsExtern = 0; + unsigned Type = 0; + + if (Target.isAbsolute()) { // constant + // SymbolNum of 0 indicates the absolute section. + // + // FIXME: Currently, these are never generated (see code below). I cannot + // find a case where they are actually emitted. + Type = macho::RIT_Vanilla; + } else { + // Resolve constant variables. + if (SD->getSymbol().isVariable()) { + int64_t Res; + if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute( + Res, Layout, Writer->getSectionAddressMap())) { + FixedValue = Res; + return; + } + } + + // Check whether we need an external or internal relocation. + if (Writer->doesSymbolRequireExternRelocation(SD)) { + IsExtern = 1; + Index = SD->getIndex(); + // For external relocations, make sure to offset the fixup value to + // compensate for the addend of the symbol address, if it was + // undefined. This occurs with weak definitions, for example. + if (!SD->Symbol->isUndefined()) + FixedValue -= Layout.getSymbolOffset(SD); + } else { + // The index is the section ordinal (1-based). + const MCSectionData &SymSD = Asm.getSectionData( + SD->getSymbol().getSection()); + Index = SymSD.getOrdinal() + 1; + FixedValue += Writer->getSectionAddress(&SymSD); + } + if (IsPCRel) + FixedValue -= Writer->getSectionAddress(Fragment->getParent()); + + Type = macho::RIT_Vanilla; + } + + // struct relocation_info (8 bytes) + macho::RelocationEntry MRE; + MRE.Word0 = FixupOffset; + MRE.Word1 = ((Index << 0) | + (IsPCRel << 24) | + (Log2Size << 25) | + (IsExtern << 27) | + (Type << 28)); + Writer->addRelocation(Fragment->getParent(), MRE); +} + +MCObjectWriter *llvm::createX86MachObjectWriter(raw_ostream &OS, + bool Is64Bit, + uint32_t CPUType, + uint32_t CPUSubtype) { + return createMachObjectWriter(new X86MachObjectWriter(Is64Bit, + CPUType, + CPUSubtype), + OS, /*IsLittleEndian=*/true); +} |
